AI accelerates Anthropic model development by almost two times, according to study

Artificial intelligence (AI) may be driving its own progress. An independent examination conducted by the non-profit organization Model Evaluation and Threat Research (METR) indicated that models developed by Anthropic are progressing almost twice as fast as before. This is partly due to the company's own use of AI systems to create more powerful versions of its technology.

The study, published this Thursday (5th), calculates that the advancement in the capabilities of Anthropic's models in specific tasks increased at a rate of 1.9 times since October 2025, compared to the previous pace.

According to METR, approximately 80% of this current acceleration is linked to the application of AI in the company's research and development activities. However, the estimate has a high degree of uncertainty.

The evaluation was carried out over an 11-week period, covering August to October 2026. Researchers had access to Anthropic's internal data, interviewed employees, and analyzed information on the performance of models that have not yet been made public.

The organization also found that the increase in computational capacity used for training does not justify the observed acceleration. The study establishes a crucial distinction: most of the progress still derives from the assistance provided by AI to experts, not from the autonomous execution of the entire development cycle.

METR stated that it found no evidence that current systems can offer highly relevant and fully independent contributions in the most critical phases of model training. Despite this, researchers detected signs that such participation may become more prominent in the future.

The advance noted by METR deserves attention because it is connected to the concept of recursive self-improvement, known by the acronym RSI. This idea posits that AI systems can increasingly contribute to the generation of new model versions. By assisting in the development of their successors with more competent systems, the process tends to gradually accelerate.

Although this scenario has not yet been achieved in a completely autonomous manner, it raises important questions about the speed of technological evolution. One concern highlighted by METR is the potential for this acceleration to become cumulative, where more advanced models assist in developing even more capable systems, which in turn would intensify the process.

The organization also did not identify clear impediments related to computing power, data availability, or experiments that would certainly block new advances in the next twelve months. However, the researchers emphasized that this analysis is also subject to uncertainties.

Despite the findings, METR recognizes significant limitations. The 1.9-fold acceleration was calculated based on performance indicators in well-defined tasks, which can be assessed through automated tests or language models. This does not imply that all aspects of AI are evolving at the same pace.

Additionally, Anthropic had the prerogative to remove confidential information from the report. According to METR, much of the evidence considered most pertinent to its conclusions was omitted, which complicates an independent assessment of all details.

The organization also highlights the uncertainty about what portion of the acceleration can be directly attributed to the use of AI, especially regarding the progress recorded at the end of 2025.

The results emerge at a time when Dario Amodei, CEO of Anthropic, advocates for prudence in the development of increasingly sophisticated systems. In September, the executive called on the sector to slow down the pace of technological capability advancement so that safeguards could keep up with the pace of improvements.

The study itself points out that understanding the speed of this evolution is vital for risk assessment. If systems begin to collaborate in a progressively more significant way in the development of their replacements, monitoring and control work may also become more complex.

Similar stories

olhardigital.com.brNovember 3, 15:21

Anthropic claims AI can accelerate its own development through recursive self-improvement

Anthropic disclosed in an update to its S-1/A prospectus, submitted to the U.S. Securities and Exchange Commission (SEC), that it may be close to creating an artificial intelligence capable of accelerating its own enhancement.

Although the company states it has not yet reached this phase, its projections suggest this could happen in the near future. This technology, known as Recursive Self-Improvement (RSI), describes a point where AI systems can improve their own development processes, boosting the evolution of their capabilities.

This perspective generates divergence: critics warn that such advances could result in extremely powerful, human-out-of-control systems capable of causing harm while pursuing their objectives. On the other hand, enthusiasts see RSI as a chance to increase productivity and speed up scientific discoveries.

Anthropic acknowledges both views. In the document sent to the SEC, the company mentions that it already uses AI in its R&D activities but warns that this approach carries a significant safety risk, especially if the pace of progress is drastically accelerated.

The prospectus update introduces new warnings in the risk factors section. Anthropic revealed that it has already used AI systems to create new models without realizing at the time the magnitude of the capacity gains achieved. The increased use of AI in research may also cause models to evolve faster than the company's ability to implement safety, alignment, and protection measures.

According to the document, internal estimates and forecasts for future models indicate that the company may soon reach what its Responsible Scaling Policy calls 'dramatically accelerated AI-attributable progress.' However, the company emphasizes that none of its current models have been classified as having exceeded this limit, and data evaluation is ongoing.

Despite this possibility, Anthropic declares that it does not foresee using Claude to autonomously design and build successive generations of advanced models without relevant human supervision or understanding. The company also plans to revise its Responsible Scaling Policy to ensure humans maintain control over the development of its most sophisticated models.

It is important to note that the filings with the SEC fulfill disclosure obligations but also serve as a means for the company to present its expectations to the market. Anthropic stresses that Claude's capabilities have improved this year, notably in programming, mathematics, and cybersecurity, although no current model has been considered by the company to have reached the stage of recursive self-improvement.

business-standard.comOctober 26, 21:01

Indian AI Startups Consider Relocation Costs to the US in Pursuit of American Capital

Despite venture capital funds creating opportunities for Indian artificial intelligence (AI) startups to enter the US market, securing initial clients in America remains a significant challenge. American companies typically require proven work experience before entering into agreements with new suppliers, which complicates startups' efforts to attract initial customers.

Venture firms such as Antler India and All In Capital have launched specialized programs to help local AI startups establish themselves in the United States. While some companies have already relocated their headquarters, found clients, and attracted subsequent funding, their expansion plans are complicated by high costs for local hiring and visa issues.

Details on Challenges

Ashish Bhatia, founder and CEO of India Accelerator, noted that winning the first few clients is one of the main difficulties for Indian startups aiming to enter the US market. He explained that firstly, buyers in the US require references, making the acquisition of the first two or three clients the most difficult. Secondly, requirements for security, regulatory compliance, and architecture are significantly higher. Thirdly, hiring sales and business development staff in the US is expensive, and visas for founders, who often move between countries, create additional obstacles.

Similarly, All In Capital identified key challenges as client access, achieving product-market fit in a new market, and building a local network. Antler India pointed out another barrier—the lack of a peer network. Founders entering a new market may not have reliable contacts for advice on hiring, pricing, legal structures, or pitching to investors.

To overcome these difficulties, venture funds provide support that goes beyond mere financing. They assist startups in finding potential clients, building local connections, and managing operational requirements.

Examples of Support Programs

Antler India launched the Embark program in May 2025. Over this period, more than 50 startups participated across four cohorts, including over 15 from India. Raghav Goyal, Investment Director at Antler India, reported that seventy-five percent of participants now have paying clients or active pilot projects in the US. More than half of them have secured subsequent rounds of funding. Nearly 50 percent have moved to the US, converting their holding company into an American structure and obtaining O-1 visas for founders. Most of them signed their first contract with an American client within three to four weeks of participating in Embark. Among the startups that relocated their headquarters to the US are Dodge, KuboCare, Nudge, and Aina.

Likewise, All In Capital launched the Golden Ticket program in January of this year to aid Indian startups in expanding into the American market. The program received over 750 applications in the first ten days and aimed to select 20–30 startups. The program's portfolio included Vasuki AI, which develops commodity hedging solutions; Tross, which creates a data layer and application programming interface (API) for AI agents in healthcare; and Mowito AI.

Among these companies, Mowito opened an office in Detroit, started pilot projects with American clients, and attracted additional funding from a Canadian investor. Kushal Bhagya, Co-founder and Partner at All In Capital, emphasized that early progress should be assessed based on the quality of client interactions and pilot projects, not just the number of users.

Recognizing the complexities of entering a new market, venture funds are increasingly offering operational support alongside capital. For instance, Antler India's Embark program facilitates introductions to American clients and provides practical assistance in early sales. It also collaborates with partners such as Deel, Rho, Alma, and Inkle, helping startups obtain visas and establish American entities.

Nevertheless, not all Indian startups aiming to expand into the US relocate their headquarters abroad. As Jasnoor Gill, Investment Director at Antler India, noted, 'Some startups keep their engineering and product teams in India while developing American teams, so relocation is more related to proximity to clients and capital than leaving India behind.'

However, the path of expansion was different for each company. Some firms participating in these programs decided to return. According to Gill, a couple of startups returned to India, usually because their core clients or teams were better served from that country. Nevertheless, the more common scenario is that founders develop higher ambitions regarding the US after spending time with American clients and investors.

technode.comOctober 23, 07:26

Startup PeakVision AI offers an alternative approach to implementing artificial intelligence in the corporate sector

For companies implementing artificial intelligence, key questions are not only about choosing a model but also about where the AI operates, its cost, and the ability to keep confidential data under the company's control. PeakVision AI offers a solution that involves moving most of the AI tasks inside the enterprise.

The company's system considers the task, rather than an individual application, as the main unit of work: users state a goal, and AI agents perform the necessary steps in a local computing environment. Co-founder and CTO Liangyu Nie told TechNode that PeakVision AI was founded in 2024, overcoming two major obstacles to AI adoption in corporations: the complexity of technology deployment and concerns over data security.

In March 2026, the company opened its headquarters in Hangzhou and received government support for research and development amounting to 15 million yuan. In May, it presented its first enterprise systems and completed an initial funding round, followed by a second round worth tens of millions of yuan in August.

PeakVision AI creates an enterprise AI stack that integrates locally deployed computing, FengKan AIOS, and AI agent services. This system is designed to focus on how agents complete tasks, rather than how people move between different applications. As Nie noted, users no longer need to load different applications; they just need to tell the AI what they want via a chat window, and it will do the work.

The startup's central thesis is that enterprise software will increasingly be structured around intent and execution, rather than menus, files, and separate applications. The company plans to make FengKan AIOS open source by the end of October. PeakVision AI monetizes through hardware, subscriptions, and custom enterprise AI implementations. The company's broader goal is to transform project-specific work into universal solutions for various industries.

FengKan AIOS supports third-party hardware, while PeakVision AI takes a selective approach to its skills ecosystem. Nie clarified that users' end skills, which may contain proprietary workflows, will remain untouched. However, more generalized skills created by partners and leading engineers can be released to the market.

Currently, the company is testing its model through the FDE academy, launched by Shanghai Sixunited Intelligent Technology with support from Advanced Micro Devices (AMD). AMD provides computing hardware, technical toolchains, and resources for projects. PeakVision AI also collaborates with chip manufacturers in early product development stages to optimize costs and performance, including a partnership with ChangXin Memory Technologies to reduce memory costs.

Unlike companies such as Wonderful, which target large, often regulated organizations and provide them with AI platforms and implementation teams, PeakVision AI targets small and medium-sized businesses. Nie emphasized that these enterprises face the same problems—costs, technical complexity, and data security—but have fewer resources to solve them. PeakVision AI uses inexpensive hardware from the Chinese supply chain so that each implementation becomes a repeatable solution for similar clients.

Nie stated: